Improved Radix-Based Approximate Homomorphic Encryption for Large Integers via Lightweight Bootstrapped Digit Carry

Homomorphic encryption (HE) for high-precision integers has been steadily researched through various schemes; however, these approaches incurred severe overhead as the bit-width grew, requiring larger parameters to support integers of several hundred to a thousand bits. A significant breakthrough was recently made by Boneh and kim (Crypto'25). Their scheme constructs a residue number system from the different slots of a single CKKS ciphertext. This enables arithmetic on thousand-bit integers without increasing parameters. However, RNS approach in Boneh et al., which performs approximate reduction, fundamentally cannot support non-arithmetic operations. Alternatively, radix-based approach proposed by Kim (CHES'25) can perform non-arithmetic operations, but they require $O(k)$ bootstraps for a bit-width $k$. This makes them highly inefficient, and thus impractical, for non-arithmetic operations requiring thousand-bit precision. This paper proposes an improved radix-based CKKS scheme, centered on a 2-step algorithm that optimizes the number of bootstraps required for the digit carry operation to $O(\log k)$. The proposed scheme requires only 3-6 bootstraps to restore the result of a 32-2048 bit integer multiplication to its unique representation, which enables the efficient implementation of non-arithmetic operations such as comparison. Furthermore, our scheme extends the radix-based system, previously limited to prime-power moduli, to support an efficient homomorphic reduction algorithm for arbitrary moduli. Furthermore, our experiments demonstrate substantial efficiency gains compared to Boneh et al. For example, for moduli used in homomorphic signatures (Curve25519, P-384, and 2048-bit RSA), our scheme can process up to 4$\times$ more integers in a single ciphertext. Specifically for Curve25519, we also reduce the latency by 1.4$\times$, shortening the amortized time by 5.6$\times$ compared to Boneh et. al. and achieving a final processing time of 1.34 seconds per data point.